Details
-
Improvement
-
Resolution: Unresolved
-
Minor
-
None
-
None
-
3
-
9223372036854775807
Description
llapi_hsm_import (really, llapi_file_open_param) performs a mount check on each and every call to make sure the path is within a Lustre filesystem. This is needless – the following IOCTL will fail if it was to a non-Lustre filesystem on import. Moreover, since some users (such as ourselves at MSFT) are in control of the time at which the importing process is launched, it may be reasonable to expect it to only come up after a successful mount of the Lustre client. Accordingly, we can elide this check. FID is also returned as an out variable and internally is translated via path2fid, but no part of our importer aside from logging uses the FID (or is expected to use it). Finally, it is desirable on import to be able to specify a source path from which to derive layout, such that files are imported with (in the most typical case) the default stripe setting set at the Lustre root. This avoids some odd OOS issues that can result by using raid0 or a statically set stripe setting on import for varying cluster sizes.
The improved import routine bumps performance from roughly ~2300 to ~4000 imports per second for a single host.
Patch to be sent shortly.